Mistral released Timestral on October 15, introducing a family of models for forecasting data that changes over time.
The four versions range from five million to 209 million parameters. Their intended inputs include sensor readings, energy demand and business activity—settings where the useful output is a numerical forecast with uncertainty around it.
Learning relationships between signals
Timestral supports multivariate forecasting, allowing related measurements to be modeled together. It can also incorporate additional information available in the past or known ahead of the forecast period.
Consider electricity demand at a factory. Historical consumption offers a starting point, while a planned production shift provides information about what will change tomorrow. Treating that schedule as an input can help distinguish an expected increase from an unexpected deviation.
This illustrates the forecasting task; it is not a reported deployment result.
The model produces 21 quantiles for each forecast step. These describe different points in the predicted distribution, giving users more information than a single central estimate.
Smaller models, practical deployment choices
Mistral reports strong results on GIFT-Eval, TIME and fev-bench, with comparisons focused on pretrained models evaluated without task-specific training.
The official repository provides standalone inference on CPUs or GPUs and supports a context of up to 32,768 time steps. These are observations in a sequence, rather than language-model tokens; their calendar coverage depends on how frequently data is recorded.
A choice between model sizes can therefore reflect operating needs as well as accuracy. Frequent forecasts across many data streams may place a different value on latency than an occasional planning exercise.
Forecasting is one part of the decision
Mistral also describes adaptation through Forge. The inference code carries an Apache 2.0 license, while model weights use a separate Modified MIT license.
AIYXL Perspective: A forecast becomes useful through the action it informs. An overstated demand estimate may create excess inventory; an understated one may leave capacity short. Evaluation should reflect those unequal costs, compare against existing methods and check whether predicted uncertainty matches observed outcomes.
The next question for Timestral is how well its general forecasting ability transfers to data that is incomplete, changes over time or behaves differently from public benchmarks.